CE 186, OCTOBER 2015
1
Automated Demand Response Refrigerator Project
Jessica Tran1 , Jordan Gilles1 , Ryan Mann2 , and Vishnu Murthy2
1
Civil and Environmental Engineering, University of California, Berkeley
2 Energy Engineering, University of California, Berkeley
Abstract – This project aimed to create a smart refrigerator for both perishables and non-perishables that is capable of using
prior usage data and the user’s choice of objectives to optimize its energy usage and performance. Using prior usage data, the
refrigerator can predict times at which it is most likely to be opened and its contents consumed, and (when in non-perishables
mode) adjust its temperature setpoint accordingly to ensure that it reaches the desired temperature at that time. In addition, an
online dashboard allows the user to view sensor data and control the refrigerator mode remotely. In either of the two modes, the
refrigerator’s compressor schedule is optimized using a thermal-model-based mixed integer linear program to minimize marginal
carbon emissions (using the WattTime API) and electricity cost (using a time-of-use rate schedule).
I. I NTRODUCTION
A. Motivation & Background
ALIFORNIA, as well as many other parts of the world,
is rapidly increasing the percentage of renewables on its
grid. However, there are significant challenges associated with
accommodating a high penetration of variable and intermittent
energy generators.
The current approach relies primarily on using natural gas
peaker plants; however, this strategy is ultimately incompatible
with long-term climate policy goals, and could result in
increased electricity prices.
Another method of reducing variability involves promoting
technological and geographic diversity among renewable generators. While effective, this tactic is not sufficient to guarantee
that power can be provided reliably.
Deployment of energy storage devices has been proposed
as an alternative strategy. Although storage is promising, cost
and scale remain as hurdles. Fortunately, there exists another
approach that is functionally equivalent to deploying massive
amounts of storage infrastructure, yet only requires the use
of relatively inexpensive electronics and software: automated
demand response.
Demand response involves strategically shifting demand for
electric power away from times when it stresses the grid and
towards times when it can be more easily accommodated.
Once primarily performed by large industrial and commercial
consumers on hot summer days after a phone call from
the local utility, DR now has the potential for adoption on
a more distributed scale with the introduction of Internetconnected home and commercial appliances. Of these, the
most promising points of use are thermostatically controlled
loads (TCLs) such as HVAC systems, refrigerators/freezers,
and hot water heaters.
This project focuses specifically on refrigerators and freezers, and on combining sensor data, user control, and real-time
marginal emissions data from the WattTime API into a modelbased optimal control framework that can provide services to
the grid while maintaining food safety. One key goal is to
ensure that the required hardware is as low-cost, modular, and
easy-to-install as possible, with the hope of future deployment
at large scale in the UC Berkeley residence halls.
C
B. Relevant Literature
The California Independent System Operator offers a useful
overview [1] of the challenges associated with having a higher
percentage of renewables on the grid.
Mathieu, Dyson, and Callaway’s 2012 ACEEE paper [2]
details how devices such as refrigerators, AC units, and
water heaters can function as a form of energy storage, and
provide much-needed services to renewables-heavy electric
power systems at low cost.
Unlike many papers involving the use of refrigerators for demand response, which make use of typical thermal resistance
and capacitance coefficients in their model, Taneja, Culler, and
Dutta’s 2010 IEEE paper [3] uses a sensor-based approach to
develop a thermal model of their refrigerator.
Another similar sensor-based thermal model is proposed
in Mann, Ju, Rosa, and Barido 2015 [4], including a simple
model of temperature changes in response to a door-opening
event.
DeWitt and Roeschke 2015 [5] includes a large amount of
relevant and helpful information about refrigerator parameter
and state estimation using sensor data, user behavior prediction, and integration of WattTime and retail rate structure
information to perform emissions and cost optimization.
C. Focus of this Study
The objective of this project is to reduce the environmental
and economic cost of operating a dorm room refrigerator.
To achieve this goal, the refrigerator will incorporate all
aspects of a cyber-physical system: gathering sensor data
from the environment, processing the data to optimize system
behavior, visualizing the data through a website dashboard,
and ultimately actuating the refrigerator compressor.
The end result is an energy consumption schedule that
minimizes electricity cost and carbon emissions while ensuring
that perishables do not expire, and non-perishables are cold
when consumed.
CE 186, OCTOBER 2015
II. T ECHNICAL D ESCRIPTION
A. Hardware
The hardware for this cyber-physical system is listed
below. The communications connectivity for these devices is
explained in the following section.
Bill of Materials:
• Small refrigerator
• Arduino Uno
• Arduino Data Logger Shield
• WiFi-enabled Raspberry Pi
• Edimax Nano USB WiFi Adapter
• Two temperature sensors (ambient and internal)
• Magnetic door position switch
• Solid-state relay to control compressor
B. Connectivity
1) Cyber-Physical System Architecture
To make these modes of operation possible and thereby minimize cost and emissions, the cyber-physical system includes
several points of data acquisition and processing:
• Arduino
– Collects data every 60 seconds from the three refrigerator sensors: a temperature sensor inside the fridge,
an ambient temperature sensor outside the fridge, and
a magnetic switch that records door position. The
sensor data is logged to an SD card on the shield,
and is sent to a Raspberry Pi over the serial port.
– Receives data about the state of the compressor from
the Raspberry Pi over the serial port, and actuates the
compressor accordingly.
• Raspberry Pi
– Receives sensor data from the Arduino through the
serial port. Parses the data into four data streams:
refrigerator temperature, ambient temperature, door
state, and measurement timestamp.
– Using an Edimax Nano USB WiFi adaptor, the
Raspberry Pi connects to the Internet and sends
each data stream to the server, hosted at netfridgejgilles.c9users.io.
– Pulls data about the compressor state from the server
and sends it to the Arduino over the serial port.
• Server
– Written in Python and hosted on Cloud9.
– Accepts sensor data from Raspberry Pi and compressor actuation command from optimization algorithm.
Records data to data streams in a SQLite database.
– Posts data from SQLite to the web page.
• Web Page
– Written in HTML, CSS, and JavaScript. Receives
data from the server.
– Shows temperature, door state, compressor state,
and refrigerator mode as exportable tables and liveupdating graphs on a dashboard interface.
– Allows user input to select the fridge mode. This
data is then sent to the server as a POST command.
2
•
Optimization Algorithm
– Written in Python, hosted locally on an Internetconnected computer.
– Makes GET request to pull most recent refrigerator
temperature, ambient temperature, compressor state,
and refrigerator mode datapoints from server.
– Pulls emissions and ambient temperature data
(current and forecasted) from the WattTime and
WeatherUnderground APIs respectively.
– Behavioral door-opening patterns, thermal model parameters, and PG&E rate structures are hard-coded
in to the script.
– Using a mixed-integer linear program, creates optimal compressor schedule for 24-hour time horizon
that minimizes electricity cost and emissions subject
to a model-based temperature forecast, temperature
constraints corresponding to the refrigerator mode,
and the requirement that the compressor cycle frequency be no less than 5 minutes.
– Send first compressor decision variable value to
server. This corresponds to the optimal compressor
state for the next 5 minutes.
– Algorithm updates with new sensor data and re-runs
optimization every 5 minutes.
The communication network between all of these devices
can be seen in Figure 1 below.
Fig. 1. Cyber-Physical System Architecture.
2) Refrigerator Modes
The refrigerator has three modes, each one representing a
common usage of a dorm fridge, which the user can select
through a web page. These modes impact the temperature
constraints within the optimization algorithm, allowing it more
or less flexibility in reducing the cost and carbon emissions of
the fridge’s performance:
• Perishable Mode: In this mode, the temperature of the
fridge must always be kept within a tight hysteretic band
(34°F - 40 °F) to ensure that the contents do not go bad.
Despite the small temperature range, it is still possible
to modify the compressor’s behavior to minimize carbon
emissions and cost.
• Drink/Non-Perishable Mode: In this mode, the fridge
has a larger temperature range (34°F - 59 °F) during
times when the refrigerator is unlikely to be opened,
and ensures that contents will be cold (34°F - 40 °F)
when its contents are most likely to be consumed. Prior
door-opening patterns were recorded and analyzed to
create a predictive model that dictates the allowable fridge
temperature range at any given time of the day. Then, the
CE 186, OCTOBER 2015
3
optimization algorithm is used to actuate the compressor
only when necessary, and in a way that minimizes carbon
emissions and cost.
• Empty Mode: In this mode, the compressor remains
off. In a residence-hall setting, this feature alone could
result in significant energy savings, and the computerbased control app provides convenience for the user.
A decision flowchart for these three modes within the
context of the cyber-physical system architecture can be seen
in Figure 2 below.
opening the door of the fridge one month ago is counted as
only 0.042 towards the score.
The score was then distributed over a 2-hour period using
the following equation:
distScore(i) = 0.25 · score(i − 3)+
(2)
0.5 · score(i − 2) + 0.75 · score(i − 1) + score(i)+
0.75 · score(i + 1) + 0.5 · score(i + 2) + 0.25 · score(i + 3)
i is the index of each time period in score.
The maximum score possible for any time period,
maxScore, corresponding to the fridge being opened at all
prior timesteps, is calculated as follows for α <1.
Z
0
maxScore =
−∞
αx dx = −
1
ln(α)
(3)
The distributed score is normalized by dividing by
maxScore such that:
distScore
(4)
maxScore
A threshold for an acceptable level of score for which the
fridge should be cold is calculated below and corresponds to
the fridge being opened about 3 times over the past week.
normScore =
Fig. 2. Refrigerator Modes Decision Flowchart.
C. Data Analysis
Analysis of two weeks of door-opening and temperature
data collected from an in-use dormitory fridge was used to
create a statistical model of user consumption behavior, and
to estimate parameter values for the thermal model used in
the optimization algorithm.
1) Behavior Prediction
When in non-perishables mode, the fridge uses behavior
prediction to ensure that the fridge is cold only when the user
is most likely to take drinks out of the fridge. Using the sample
data, an algorithm predicts when the fridge is likely to be
opened.
First, the timestamps of all refrigerator-opening events are
recorded to an array openT ime. For simplicity, a current time
currT ime is set to be a few hours after the most recent
opening of the fridge. The age of each openT ime in days
was recorded by comparing currT ime with openT ime, and
was recorded to an array openAge.
Each openT ime was analyzed to determine when in a 15minute discretized 24-hour period the door was opened. This
value was then added to a score array using the following
equation:
weightedScore(i) = score(i) + αage(j)
(1)
i is an integer between 0 and 95, representing a 15-minute
time period within the 24-hour day.
j is the index of each value in the age array.
For this analysis, the decay factor α is set to 0.9. This
factor places a large emphasis on recent door-opening events
and a smaller emphasis on more distant events. For example,
α7 + α5 + α3
(5)
maxScore
The distributed score is then compared to the threshold. If
the distributed score distScore is larger than the threshold,
the fridge will be constrained to tighter temperature constraints. Otherwise, it will be constrained to the wider temperature bounds.
The distScore is plotted in the histogram Figure 3 below,
and the threshold is shown as a horizontal dotted line.
threshold =
Fig. 3. Door Openings Histogram.
As an important note, although this analysis was done for
two weeks of collected data, the analysis is set up such that
it can be updated in real time with each fridge opening. This
would allow the system to learn over time when the fridge is
likely to be used and, therefore, should be cold.
CE 186, OCTOBER 2015
4
2) Thermal Model Parameter Estimation
Data was collected over a two-week period using an
Arduino in a Berkeley dorm fridge. The fridge was the same
model as was used for this project, and was utilized by the
room’s three occupants. Just as in the final system, fridge
temperature, outside temperature, and door state was collected
from the fridge. After the two-week period, the data was
analyzed to determine when the compressor was on, but also
when the user was most likely to use the fridge.
A simple algorithm was used to determine when the fridge
was on; if the temperature in the fridge was decreasing, it
was assumed that the compressor was on; otherwise, it was
assumed that the compressor was off. This analysis revealed
that the compressor simply turned on at regular intervals
regardless of the current temperature inside the fridge.
The fridge temperature, outside temperature, and compressor state was sent into an algorithm to determine parameters
that could be used to predict future temperatures of the fridge.
A linear equation of the following form was used:
Tf (k + 1) = θ1 Tf (k) + θ2 Ta (k) + θ3 s(k)
(6)
Tf (k) is the temperature of the refrigerator at timestep k,
Ta (k) is the ambient temperature of the refrigerator at timestep
k, and s(k) is the on/off state of the compressor. θ1 , θ2 , and
θ3 are parameters corresponding to the thermal properties of
the refrigerator.
The first week of fridge data was processed by an online
linear regression model to determine the thermal model parameters. The data from the second week was not analyzed so
that the results of the parameter estimate could be tested.
For the test, the parameters from the regression analysis
were used to predict the temperature of the fridge over the
second week, given the initial refrigerator temperature, the
temperature in the room, and the compressor state. The results
are plotted below in Figure 4:
Fig. 4. Modeled vs. Measured Refrigerator Temperatures.
As the plot shows, the linear model closely matched the
measured temperature values. Given only the starting temperature of the fridge, over a one week period, the predicted
temperature only varied from the actual temperature by a few
degrees.
3) Optimization Algorithm
The results of the behavior prediction and thermal parameter
estimation, along with PG&E rate data, were hard-coded
into the optimization algorithm. Current refrigerator mode
and temperature were pulled from the server, and marginal
emissions and ambient temperature predictions were pulled
from the WattTime and Weather Underground APIs.
The optimization algorithm used is a slight variation on
the mixed-integer linear program developed in DeWitt and
Roeschke 2015 [5]:
N
−1
X
(λc(k) + (1 − λ)e(k))P s(k)
(7)
Tf (k + 1) = θ1 Tf (k) + θ2 Ta (k) + θ3 s(k)
(8)
Tf,min,on ≤ Tf (i) ≤ Tf,max,on
(9)
min
s(k),Tf (k)
k=0
Subject to:
Tf,min,of f ≤ Tf (j) ≤ Tf,max,of f
(10)
Tf (0) = Tf,o
(11)
0 ≤ s(k − 5) + s(k − 4) + s(k − 3)
(12)
+ s(k − 2) − 4s(k − 1) + 5s(k) ≤ 5
s(k) = [0, 1]
(13)
i∈k=
b [7 : 15 − 9 : 00] ∪ [10 : 00 − 11 : 45]∪
[16 : 30 − 17 : 30] ∪ [21 : 45 − 23 : 15]
(14)
j∈k=
b [0 : 00 − 7 : 15] ∪ [9 : 00 − 10 : 00]∪
(15)
[11 : 45 − 16 : 30] ∪ [17 : 30 − 21 : 45] ∪ [23 : 15 − 24 : 00]
λ is a multi-objective optimization constant that determines
the relative weights of the electricity and carbon cost functions.
For this project, a value of 0.6 was used, giving carbon
intensity a slightly higher influence over the end result. This
stems from the lack of significant variability in retail electricity
rates, and any potential mismatch between predetermined
electricity rates and real-time grid conditions.
c(k) and e(k) are the carbon and electricity cost functions
respectively. PG&E’s E-20 summer time-of-use rates for large
commercial customers was used, as this appears to be the rate
paid by the University of California for electricity. P is the
energy consumption of the compressor, assumed to be 0.1 kW
when in use.
The refrigerator temperature Tf changes in time according
to a linear model using the estimated parameter values. The
refrigerator temperature must stay within the set constraints
for each timestep. The initial modeled temperature is equal
to the most recent sensor value pulled from the server. The
compressor is not allowed to cycle more frequently than once
per 5 minutes.
i represents times when the refrigerator is likely to be
opened. The refrigerator temperature is constrained to the
range [34°F - 40 °F], for both the perishable and nonperishable refrigerator modes. j represents times when the
refrigerator is not likely to be opened. The refrigerator temperature is constrained to the range [34°F - 59 °F] when the
refrigerator is in non-perishable mode.
CE 186, OCTOBER 2015
5
D. Visualization
The first step of the visualization is logging on to a personal
fridge website, seen in Figure 5. If deployed at scale, one login
and password would be assigned to each room, and would need
to be stored in a separate database.
Fig. 7. Refrigerator Mode Selection Notification.
Fig. 5. Web Dashboard Login Page.
This data is then sent to the server. This is done through a
POST request run by jQuery and Ajax. The selection can then
be read by the optimization algorithm.
Additionally, the Analytics page (Figure 8) is designed to
show a histogram describing historic usage patterns for each
individual fridge. This data is an important parameter used in
the optimization algorithm.
Once the user has logged on, the website redirects to a
dashboard interface, which provides the user with an overview
of the site, along with a sidebar containing further details about
the user’s refrigerator. Perhaps the most significant of these is
a real-time data stream found on the Plots page. Figures 6
shows a sample plot of interior temperature data.
Fig. 8. Analytics Page.
All parts of the visualization were created using HTML,
CSS, Javascript, and jQuery. Aside from the images shown,
the website also includes an Export sub-menu which displays
a table containing the most recent data points.
Fig. 6. Refrigerator Temperature Plot.
The dashboard features radio buttons corresponding to the
three refrigerator modes: Perishables, Drinks, and Off. As
previously described, these three different options were chosen
to most accurately reflect the variety of contents that could be
placed in a dorm fridge. When one of those modes is clicked
by the user, a notification will pop up on the present browser
confirming the selection, as seen in Figure 7.
CE 186, OCTOBER 2015
6
III. D ISCUSSION
IV. S UMMARY
The original motivation for this cyber-physical system
project came from the authors’ personal experience with
dormitory refrigerators in their freshman and sophomore years.
Although dorm fridges do prevent perishables from expiring,
their unsophisticated, timer-based compressor scheduling controls mean that they consume far more energy than their role
requires. By contrast, the NetFridge developed in this project
uses data from WattTime API, PG&E utility rates, prior fridge
usage data, real-time fridge sensors, and user preferences to
optimize how to best operate the fridge.
Although dorm fridges are set to store perishable food 24/7,
this rarely matches the actual fridge contents, as students living
in the dorms generally only have perishable food when there
are leftovers from the dining hall. The rest of the time, the
fridge is storing drinks, or nothing at all.
NetFridge’s combination of behavior prediction, user mode
selection, and compressor optimization allow for significant
energy reduction, while preserving core food-safety functionality.
Although significant progress was made this semester, there
are several opportunities for further improvement and study
that could be implemented by future CE 186 students:
• Install the NetFridge for an extended period of time
in actual dorm conditions, and compare energy savings,
carbon emissions and demand reduction to a control
group of non-optimally-actuated refrigerators.
• Incorporate real-time locational marginal (wholesale)
pricing as an alternate cost function, and to estimate the
grid infrastructure benefits of smart fridge.
• Improve ambient indoor temperature forecasting by using
previous days building temperature data, ambient temperature sensor readings, and the Weather Underground
API’s forecast.
• Use Kalman Filter state-estimation algorithm to recompute the refrigerator model parameters every time
the refrigerator is opened. Sensor data suggests that the
thermal capacitance changes significantly when food or
drinks are added or removed from the refrigerator.
• Update optimization bounds in real time with stream of
past behavior data.
• Host optimization on the server, instead of running it
locally on a computer.
• Replace Arduino and Raspberry Pi with lower-cost hardware to improve scalability.
• Create an enclosure for all sensors and electronics that
protects them from damage while allowing for easy
installation.
• Allow users to set future temperature setpoints, overriding
or augmenting the optimization.
The automated-demand-response refrigerator incorporates
all features of a cyber-physical system as follows:
1) Infrastructure: The fridge interacts with California’s
electric grid through the minimization of peak demand
and carbon emissions.
2) Sensing: The cyber-fridge senses the conditions of the
fridge through two temperature sensors, a door position
switch, and the compressor state. In addition, the fridge
uses past behavior to ensure that non-perishables are
cold when consumed.
3) Actuation and Decision Support: The main point of
actuation in the system is the compressor on the fridge,
which turns off or on depending on refrigerator state
and grid conditions. There is also user decision support
in the data displayed to the user through the web page,
informing the user’s choice of fridge mode.
4) Connectivity: The cyber-fridge is connected to the user
through the web app, is informed about the state of the
grid through WattTime data, and impacts the grid by
providing demand response services.
5) Data Analysis: The majority of the data analysis happens in the Python processing stage. Here, all of the
data streams are integrated and analyzed to determine
whether the compressor should be on or off.
An estimate provided in Dewitt and Roeshke 2015 [5]
suggests that NetFridge may be able to provide up to 68%
electricity savings per dorm room, assuming that is is run
in non-perishables mode for 6 hours per day on average.
Each fridge uses approximately 200 kWh annually; if each
of the 1,700 dormitories at UC Berkeley were outfitted with a
NetFridge, total energy savings could amount to approximately
271,660 kWh. Using an average electricity cost of $0.10/kWh,
total annual savings for the campus aggregate to $27,166.
Given that this technology is scalable, implementation across
other campuses is easy to imagine, and could contribute to
further economic and environmental benefits.
CE 186, OCTOBER 2015
7
V. ACKNOWLEDGEMENTS
The authors would like to thank Professor Scott Moura
and Eric Burger for providing the hardware used in this
project, as well as their assistance with server setup and many
other aspects of the software development process. We also
would like to thank Matthew Roeshke and Zoltan DeWitt for
allowing us to make use of their optimization code, and for
helping us install the solver required to run it. In addition,
thanks to Anna Schneider from WattTime for her assistance
in accessing the WattTime API. Finally, we would like to
thank Diego Ponce de Leon Barido, Javier Rosa, and Sandro
Iacovella, whose FlexBox refrigerator sensing and control
research project inspired our own.
R EFERENCES
[1] California Independent System Operator, ’What the duck curve tells us
about managing a green grid’, 2013.
[2] J. Mathieu, M. Dyson, and D. Callaway, ’Using Residential Electric Loads
for Fast Demand Response: The Potential Resource and Revenues, the
Costs, and Policy Recommendations’, American Council for an EnergyEfficient Economy, 2012.
[3] J. Taneja, D. Culler, and P. Dutta, ’Towards Cooperative Grids: Sensor/Actuator Networks for Renewables Integration’, Institute of Electrical
and Electronics Engineers, 2010.
[4] R. Mann, J. Ju, J. Rosa, and D. Barido, ’Sensor-based validation of
thermostatically controlled load refrigerator models’, Energy Modeling,
Analysis, and Control Group, 2015.
[5] Z. DeWitt and M. Roeschke, ’Optimal Refrigeration Control for Soda
Vending Machines’, Energy, Controls, & Applications Lab, 2015.